TL;DR
This paper investigates whether stereotypes captured in natural language text (from crowd-workers, ChatGPT, and Twitter) align with traditional psychological survey measures. The authors find that while some dimensions like Morality and Status correlate well, others like Beliefs are less consistently represented, suggesting that natural language data offers richer, more nuanced insights into stereotypes than traditional scales alone.
What Problem Does This Paper Address?
For decades, psychologists have studied stereotypes using specially designed rating scales. However, with the rise of large text corpora and NLP tools, researchers now have the opportunity to study stereotypes "in the wild." A key question is whether these new, large-scale data sources truly capture the same information as traditional methods. This paper addresses this by comparing stereotype measurements across six psychologically-motivated dimensions (Sociability, Morality, Ability, Assertiveness, Beliefs, and Status) derived from traditional survey data against those extracted from crowd-generated sentences, ChatGPT-generated sentences, and Twitter data. The goal is to understand the similarities, differences, and complementary value of these diverse data sources for stereotype research.
Methods
The core of the methodology involves using the POLAR model (adapted from Mathew et al., 2020) to transform high-dimensional sentence embeddings (from RoBERTa) into a six-dimensional space that aligns with the six stereotype dimensions. This model projects sentence representations onto vectors defined by pairs of antonymous adjectives (e.g., "friendly" vs. "unfriendly" for Sociability). For comparison, a word-counting baseline was also used for adjective lists. Additionally, BERTopic was employed for topic modeling to explore the qualitative nuances within the text data that might explain or differentiate the quantitative stereotype ratings.
Datasets
The study utilizes several datasets:
- Gold Standard Rating Scales: Data from Nicolas et al. (2022), where 400 Amazon Turk workers rated social groups on six dimensions.
- Spontaneous Adjectives: Adjectives collected alongside the rating scales in Nicolas et al. (2022).
- StereoSet: A dataset of stereotypical, anti-stereotypical, and neutral sentences generated by crowd-workers (Nadeem et al., 2020).
- ChatGPT: Novel sentences generated using ChatGPT based on specific prompts about social groups.
- Twitter: Tweets containing the substring "GROUP are" for the target groups, filtered for relevance and quality.
The study focuses on 10 occupational groups common across these datasets: Politicians, Teachers, CEOs, Scientists, Bankers, Accountants, Engineers, Farmers, Lawyers, and Nurses.
Key Insights and Ideas
- Morality and Status are Robust: These dimensions show consistently high correlations across all text data sources with traditional scales, suggesting they are reliably captured in natural language.
- Beliefs are Elusive: The Beliefs dimension is infrequently mentioned across all text datasets, making it difficult to estimate reliably.
- Ability is Occupation-Centric: The Ability dimension is highly prevalent, especially when discussing occupational groups, aligning with the nature of the target groups.
- Natural Language Adds Nuance: While text data may not perfectly replicate scale-based results, it provides richer, more specific interpretations of stereotypes (e.g., politicians being "corrupt" rather than just "low Morality").
- Differentiating Similar Groups: Text data can reveal subtle differences between groups that are rated similarly on abstract scales (e.g., different aspects of "Ability" for scientists vs. nurses).
- Twitter's Heterogeneity: The Twitter dataset shows lower correlations for Sociability, likely due to its diverse and less explicitly stereotypical nature compared to curated datasets.
Key Results
- Correlation with Scales: Morality and Status dimensions showed correlations generally above 0.5 across most text datasets. Assertiveness and Beliefs showed moderate correlations. Sociability showed varied correlations, with Twitter data exhibiting negative correlations.
- Prevalence of Dimensions: Ability was the most prevalent dimension across many occupational groups. Morality was also frequently mentioned, particularly for politicians. Beliefs were consistently the least prevalent.
- Representativeness: Text data can highlight which dimensions are most "representative" or salient for specific groups. For instance, morality is highly representative for politicians, while ability is for scientists and engineers.
- Qualitative Insights: Topic modeling revealed specific, often harmful, stereotypes not captured by abstract scales, such as scientists being stereotyped as male or nurses as exclusively female.
Limitations and Future Work
- Language and Culture: The study is limited to English-language resources and primarily North American contexts. Future work should expand to other languages and cultures.
- Group Representation: The study focused on occupational groups. Future research should investigate stereotypes related to gender, ethnicity, and socio-economic status.
- Topic Modeling Sensitivity: Topic modeling and clustering methods are sensitive to parameter settings, which can influence findings.
- Data Bias: Each data source has its own sampling bias, which may contribute to observed differences.
- Counter-Strategies: The authors suggest that understanding the strengths and weaknesses of different data sources is crucial for developing effective counter-strategies against harmful stereotypes. Future work could explore how to augment traditional sources with data from unconventional sources like social media and LLMs.
Summary updated